File size: 7,920 Bytes
9d22eee
2a5f9fb
 
df66f6e
 
9d22eee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
 
 
9d22eee
 
 
 
 
 
 
 
 
 
 
 
 
b762711
9d22eee
 
 
9b2e755
9d22eee
9b2e755
 
9d22eee
 
 
 
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b762711
b1a1395
2a5f9fb
 
 
 
 
 
 
 
460ecf2
2a5f9fb
 
b1a1395
 
 
 
 
 
 
 
 
 
 
 
 
b762711
b1a1395
 
 
 
 
 
 
 
9b2e755
b1a1395
2a5f9fb
 
9d22eee
2a5f9fb
9d22eee
2a5f9fb
 
 
9d22eee
9d6aecc
05bda40
 
193f184
9d22eee
2a5f9fb
 
 
 
 
 
 
 
9d6aecc
 
2a5f9fb
 
05bda40
 
 
 
2a5f9fb
 
9d22eee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
 
 
 
9b2e755
 
2a5f9fb
 
 
 
b1a1395
2a5f9fb
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd

def fields(raw_class):
    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]


@dataclass
class Task:
    benchmark: str
    metric: str
    col_name: str

class Tasks(Enum):
    arc = Task("arc:challenge", "acc_norm", "ARC")
    hellaswag = Task("hellaswag", "acc_norm", "HellaSwag")
    mmlu = Task("hendrycksTest", "acc", "MMLU")
    truthfulqa = Task("truthfulqa:mc", "mc2", "TruthfulQA")
    winogrande = Task("winogrande", "acc", "Winogrande")
    gsm8k = Task("gsm8k", "acc", "GSM8K")

# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    dummy: bool = False

auto_eval_column_dict = []
# Init
auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
#Scores
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
# Model information
auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
auto_eval_column_dict.append(["merged", ColumnContent, ColumnContent("Merged", "bool", False)])
auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, hidden=True)])
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
auto_eval_column_dict.append(["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
auto_eval_column_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)])

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)

@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model = ColumnContent("model", "markdown", True)
    revision = ColumnContent("revision", "str", True)
    private = ColumnContent("private", "bool", True)
    precision = ColumnContent("precision", "str", True)
    weight_type = ColumnContent("weight_type", "str", "Original")
    status = ColumnContent("status", "str", True)


baseline_row = {
    AutoEvalColumn.model.name: "<p>Baseline</p>",
    AutoEvalColumn.revision.name: "N/A",
    AutoEvalColumn.precision.name: None,
    AutoEvalColumn.merged.name: False,
    AutoEvalColumn.average.name: 31.0,
    AutoEvalColumn.arc.name: 25.0,
    AutoEvalColumn.hellaswag.name: 25.0,
    AutoEvalColumn.mmlu.name: 25.0,
    AutoEvalColumn.truthfulqa.name: 25.0,
    AutoEvalColumn.winogrande.name: 50.0,
    AutoEvalColumn.gsm8k.name: 0.21,
    AutoEvalColumn.dummy.name: "baseline",
    AutoEvalColumn.model_type.name: "",
    AutoEvalColumn.flagged.name: False,
}

# Average ⬆️ human baseline is 0.897 (source: averaging human baselines below)
# ARC human baseline is 0.80 (source: https://lab42.global/arc/)
# HellaSwag human baseline is 0.95 (source: https://deepgram.com/learn/hellaswag-llm-benchmark-guide)
# MMLU human baseline is 0.898 (source: https://openreview.net/forum?id=d7KBjmI3GmQ)
# TruthfulQA human baseline is 0.94(source: https://arxiv.org/pdf/2109.07958.pdf)
# Winogrande: https://leaderboard.allenai.org/winogrande/submissions/public
# GSM8K: paper
# Define the human baselines
human_baseline_row = {
    AutoEvalColumn.model.name: "<p>Human performance</p>",
    AutoEvalColumn.revision.name: "N/A",
    AutoEvalColumn.precision.name: None,
    AutoEvalColumn.average.name: 92.75,
    AutoEvalColumn.merged.name: False,
    AutoEvalColumn.arc.name: 80.0,
    AutoEvalColumn.hellaswag.name: 95.0,
    AutoEvalColumn.mmlu.name: 89.8,
    AutoEvalColumn.truthfulqa.name: 94.0,
    AutoEvalColumn.winogrande.name: 94.0,
    AutoEvalColumn.gsm8k.name: 100,
    AutoEvalColumn.dummy.name: "human_baseline",
    AutoEvalColumn.model_type.name: "",
    AutoEvalColumn.flagged.name: False,
}

@dataclass
class ModelDetails:
    name: str
    symbol: str = "" # emoji, only for the model type


class ModelType(Enum):
    PT = ModelDetails(name="pretrained", symbol="🟢")
    CPT = ModelDetails(name="continuously pretrained", symbol="🟩")
    FT = ModelDetails(name="fine-tuned on domain-specific datasets", symbol="🔶")
    chat = ModelDetails(name="chat models (RLHF, DPO, IFT, ...)", symbol="💬")
    merges = ModelDetails(name="base merges and moerges", symbol="🤝")
    Unknown = ModelDetails(name="", symbol="?")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        if "fine-tuned" in type or "🔶" in type:
            return ModelType.FT
        if "continously pretrained" in type or "🟩" in type:
            return ModelType.CPT
        if "pretrained" in type or "🟢" in type:
            return ModelType.PT
        if any([k in type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]):
            return ModelType.chat
        if "merge" in type or "🤝" in type:
            return ModelType.merges
        return ModelType.Unknown

class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")

class Precision(Enum):
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")
    qt_8bit = ModelDetails("8bit")
    qt_4bit = ModelDetails("4bit")
    qt_GPTQ = ModelDetails("GPTQ")
    Unknown = ModelDetails("?")

    def from_str(precision):
        if precision in ["torch.float16", "float16"]:
            return Precision.float16
        if precision in ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["8bit"]:
            return Precision.qt_8bit
        if precision in ["4bit"]:
            return Precision.qt_4bit
        if precision in ["GPTQ", "None"]:
            return Precision.qt_GPTQ
        return Precision.Unknown
        



# Column selection
COLS = [c.name for c in fields(AutoEvalColumn)]
TYPES = [c.type for c in fields(AutoEvalColumn)]

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

BENCHMARK_COLS = [t.value.col_name for t in Tasks]

NUMERIC_INTERVALS = {
    "?": pd.Interval(-1, 0, closed="right"),
    "~1.5": pd.Interval(0, 2, closed="right"),
    "~3": pd.Interval(2, 4, closed="right"),
    "~7": pd.Interval(4, 9, closed="right"),
    "~13": pd.Interval(9, 20, closed="right"),
    "~35": pd.Interval(20, 45, closed="right"),
    "~60": pd.Interval(45, 70, closed="right"),
    "70+": pd.Interval(70, 10000, closed="right"),
}